From Deterministic Code to Autonomous Behavior: Building Reliable AI Agents
Traditional software is a known quantity: a deterministic function where you control the code and predict the output. Agentic software, however, is more like a brilliant but unpredictable employee hired off the internet—it can plan, reason, and use tools, but it can also hallucinate and fail silently. This tutorial provides software architects and engineers with a concrete framework to treat trust not as an afterthought, but as a first-class architectural concern.
We begin by deconstructing the "Human-in-the-Loop" trap, exploring why human attention is a scarce resource and how typical approval flows often lead to dangerous machine deference. We then dive into the four pillars of agentic security, providing specific mitigations for the new "intelligence layer" of the attack surface. Finally, we explore advanced evaluation and observability strategies, including the use of "LLM-as-a-Judge" and conflicting-goals architectures to ensure that autonomous systems remain within their intended guardrails. Participants will leave with the tools to demonstrate trust through evidence, not just claims.